A method of action recognition

Abstract

The invention discloses an action recognition method, comprising: step 1): obtaining a picture sequence set through a device; step 2): constructing a deep learning target detection network, inputting the picture sequence set into the detection network, and obtaining a detection frame set; step 3): Convert the detection frame set into a space-time map; Step 4): Obtain the picture, device ID number, and image timestamp, and combine the space-time map to generate trajectory and compare the trajectory array; Step 5): Pass the space-time map and the trajectory array Update the trajectory array and confirm the action; Step 6): Relay update the trajectory array according to the timestamp to keep the trajectory array dynamic; the action recognition method is intelligent, efficient and accurate, and can be embedded in any scene and neural network , with a high degree of modularity.

Description

technical field [0001] The present invention relates to an action recognition method. Background technique [0002] Video understanding and recognition is one of the basic tasks of computer vision. Compared with images, video content and background are more complex and changeable. Different action categories have similarities, and the same category has different characteristics in different environments. specialty. In addition, occlusion, jitter, and angle of view changes caused by shooting also bring further difficulties to action recognition. In practical applications, accurate action recognition is helpful for public opinion monitoring, advertising placement, and many other tasks related to video understanding. With the rapid development of deep neural network technology in various fields of computer vision, the use of artificial intelligence for video-based action recognition tasks has become very common. The specific application methods are generally divided into the...

AI Technical Summary

Problems solved by technology

Two-stream two-dimensional convolution uses action recognition that models time and space separately to extract spatio-temporal features, and then weights and fuses various features through average pooling or time-series structures such as LSTM and RNN or adds attention to the mechanism. To get the final result, the specificity of this method is that the time complexity is low, but the precision is not high

[0006] The action recognition methods currently used are single, and the applicability is not high. The actual industry and application requirements are not only for action recognition, but also for a series of requirements such as target detection. The time complexity requirement is as small as possible. , it is impossible to add a set of networks because of the requirement of adding an action recognition, and most of the current methods do not make full use of the motion relationship between the frames before and after the timing, which leads to the fact that most of the existing recognition methods do not have good performance at the same time. Accuracy, less timeliness, wider robustness

Images